Computationally Generated Cardiac Biomarkers: Heart Rate Patterns to Predict Death Following Coronary Attacks

摘要

Heart disease is the leading cause of death in the United States, claiming over 830,000 lives each year (34% of all deaths or roughly one death every 38 seconds). A similar situation exists in other parts of the world, where an estimated 40% of all deaths in the developing world by the year 2020 are expected to be due to heart disease. The risk of death in these patients can be substantially lowered through the delivery of appropriate treatments (e.g., pharmacological and surgical interventions). However, matching patients to treatments that are appropriate for their risk remains challenging. In this paper, we aim to address this challenge by developing novel computational biomarkers that can be used to risk stratify patients. Our focus is on identifying high risk behavior in large datasets of electrocardiographic (ECG) signals from patients who experienced mortality following coronary attacks and those that remained event free. We frame the problem of finding risk markers as the discovery of approximately conserved heart rate sequences that are significantly over-represented in either high or low risk patients. We propose a randomized hashing- and greedy centroid selection-based algorithm to efficiently discover such heart rate patterns in large high-resolution ECG datasets captured continuously over long periods from thousands of patients. When evaluated on data from 3,067 patients in two separate cohorts, our pattern discovery algorithm was able to correctly identify patients at high risk of death, even after adjusting for information in existing heart rate-based risk stratification metrics. Moreover, our approach can be easily extended to other clinical and non-clinical applications focused on approximate sequential patterns discovery in massive time-series datasets.